In [1]:
import pandas as pd
import numpy as np
import datetime
import matplotlib.pyplot as plt
plt.style.use('seaborn')
%matplotlib inline
import warnings
warnings.filterwarnings("ignore")
In [2]:
#read the data
std = pd.read_csv('time_std.csv')
e7 = pd.read_csv('time_e7.csv')
std = std[['date','cost_mean','cost_median','cost_sum','usage']]
e7 = e7[['date','cost_mean','cost_median','cost_sum','usage']]
In [3]:
# groupby 'date', and use the mean value to analysis
std_low = std[std['usage']=='low'][['date','cost_mean']].groupby('date').mean()
std_med = std[std['usage']=='med'][['date','cost_mean']].groupby('date').mean()
std_hig = std[std['usage']=='hig'][['date','cost_mean']].groupby('date').mean()
e7_low = e7[e7['usage']=='low'][['date','cost_mean']].groupby('date').mean()
e7_med = e7[e7['usage']=='med'][['date','cost_mean']].groupby('date').mean()
e7_hig = e7[e7['usage']=='hig'][['date','cost_mean']].groupby('date').mean()
In [4]:
std_low.index = pd.to_datetime(std_low.index)
std_med.index = pd.to_datetime(std_med.index)
std_hig.index = pd.to_datetime(std_hig.index)
e7_low.index=pd.to_datetime(e7_low.index)
e7_hig.index=pd.to_datetime(e7_hig.index)
e7_med.index=pd.to_datetime(e7_med.index)
std_low.reset_index(inplace=True)
std_med.reset_index(inplace=True)
std_hig.reset_index(inplace=True)
e7_low.reset_index(inplace=True)
e7_med.reset_index(inplace=True)
e7_hig.reset_index(inplace=True)
In [6]:
# Create confidence level
std_low['low_CI'] = std_low['cost_mean'] - np.std(std_low['cost_mean'])
std_low['upper_CI'] = std_low['cost_mean'] + np.std(std_low['cost_mean'])
std_med['low_CI'] = std_med['cost_mean'] - np.std(std_med['cost_mean'])
std_med['upper_CI'] = std_med['cost_mean'] + np.std(std_med['cost_mean'])
std_hig['low_CI'] = std_hig['cost_mean'] - np.std(std_hig['cost_mean'])
std_hig['upper_CI'] = std_hig['cost_mean'] + np.std(std_hig['cost_mean'])
e7_low['low_CI'] = e7_low['cost_mean'] - np.std(e7_low['cost_mean'])
e7_low['upper_CI'] = e7_low['cost_mean'] + np.std(e7_low['cost_mean'])
e7_med['low_CI'] = e7_med['cost_mean'] - np.std(e7_med['cost_mean'])
e7_med['upper_CI'] = e7_med['cost_mean'] + np.std(e7_med['cost_mean'])
e7_hig['low_CI'] = e7_hig['cost_mean'] - np.std(e7_hig['cost_mean'])
e7_hig['upper_CI'] = e7_hig['cost_mean'] + np.std(e7_hig['cost_mean'])
std_low = std_low[std_low['date']>='20120201']
std_med = std_med[std_med['date']>='20120201']
std_hig = std_hig[std_hig['date']>='20120201']
e7_low = e7_low[e7_low['date']>='20120201']
e7_med = e7_med[e7_med['date']>='20120201']
e7_hig = e7_hig[e7_hig['date']>='20120201']
Overview¶
In [7]:
plt.figure(figsize=(15,5),dpi=300)
plt.plot(std_low['date'],std_low['cost_mean'],lw=2, alpha = 1,label='std_low')
plt.fill_between(std_low['date'], std_low['low_CI'], std_low['upper_CI'], alpha = 0.3,label='std_low_CI')
plt.plot(e7_low['date'],e7_low['cost_mean'],lw=2, alpha = 1,label='e7_low')
plt.fill_between(e7_low['date'], e7_low['low_CI'], e7_low['upper_CI'], alpha = 0.3,label='e7_low_CI')
plt.title('STD vs E7, low')
plt.legend(loc='best',fontsize=10)
plt.savefig('low.jpg')
In [10]:
plt.figure(figsize=(15,5),dpi=300)
plt.plot(std_med['date'],std_med['cost_mean'],lw=2, alpha = 1,label='std_med')
plt.fill_between(std_med['date'], std_med['low_CI'], std_med['upper_CI'], alpha = 0.3,label='std_med_CI')
plt.plot(e7_med['date'],e7_med['cost_mean'],lw=2, alpha = 1,label='e7_med')
plt.fill_between(e7_med['date'], e7_med['low_CI'], e7_med['upper_CI'], alpha = 0.3,label='e7_med_CI')
plt.title('STD vs E7, med')
plt.legend(loc='best',fontsize=10)
plt.savefig('med.jpg')
In [9]:
plt.figure(figsize=(15,5),dpi=300)
plt.plot(std_hig['date'],std_hig['cost_mean'],lw=2, alpha = 1,label='std_hig')
plt.fill_between(std_hig['date'], std_hig['low_CI'], std_hig['upper_CI'], alpha = 0.3,label='std_hig_CI')
plt.plot(e7_hig['date'],e7_hig['cost_mean'],lw=2, alpha = 1,label='e7_hig')
plt.fill_between(e7_hig['date'], e7_hig['low_CI'], e7_hig['upper_CI'], alpha = 0.3,label='e7_hig_CI')
plt.title('STD vs E7, hig')
plt.legend(loc='best',fontsize=10)
plt.savefig('hig.jpg')
peak/off-peak split¶
- try different ratios to find in which conditions customers can benefit from E7
- take 0.72/0.28 to analysis
In [12]:
time_e7_72 = pd.read_csv('time_e7_72.csv')
In [13]:
e7_low_72 = time_e7_72[time_e7_72['usage']=='low'][['date','cost_mean']].groupby('date').mean()
e7_med_72 = time_e7_72[time_e7_72['usage']=='med'][['date','cost_mean']].groupby('date').mean()
e7_hig_72 = time_e7_72[time_e7_72['usage']=='hig'][['date','cost_mean']].groupby('date').mean()
In [14]:
e7_low_72.index=pd.to_datetime(e7_low_72.index)
e7_hig_72.index=pd.to_datetime(e7_hig_72.index)
e7_med_72.index=pd.to_datetime(e7_med_72.index)
In [15]:
e7_low_72.reset_index(inplace=True)
e7_med_72.reset_index(inplace=True)
e7_hig_72.reset_index(inplace=True)
#Create confidence interval
e7_low_72['low_CI'] = e7_low_72['cost_mean'] - np.std(e7_low_72['cost_mean'])
e7_low_72['upper_CI'] = e7_low_72['cost_mean'] + np.std(e7_low_72['cost_mean'])
e7_med_72['low_CI'] = e7_med_72['cost_mean'] - np.std(e7_med_72['cost_mean'])
e7_med_72['upper_CI'] = e7_med_72['cost_mean'] + np.std(e7_med_72['cost_mean'])
e7_hig_72['low_CI'] = e7_hig_72['cost_mean'] - np.std(e7_hig_72['cost_mean'])
e7_hig_72['upper_CI'] = e7_hig_72['cost_mean'] + np.std(e7_hig_72['cost_mean'])
e7_low_72 = e7_low_72[e7_low_72['date']>='20120201']
e7_med_72 = e7_med_72[e7_med_72['date']>='20120201']
e7_hig_72 = e7_hig_72[e7_hig_72['date']>='20120201']
In [16]:
plt.figure(figsize=(15,5),dpi=300)
plt.plot(std_low['date'],std_low['cost_mean'],lw=2, alpha = 1,label='std_low')
plt.fill_between(std_low['date'], std_low['low_CI'], std_low['upper_CI'], alpha = 0.3,label='std_low_CI')
plt.plot(e7_low_72['date'],e7_low_72['cost_mean'],lw=2, alpha = 1,label='e7_low')
plt.fill_between(e7_low_72['date'], e7_low_72['low_CI'], e7_low_72['upper_CI'], alpha = 0.3,label='e7_low_CI')
plt.title('STD vs E7, low72')
plt.legend(loc='best',fontsize=10)
plt.savefig('low_72.jpg')
In [17]:
plt.figure(figsize=(15,5),dpi=300)
plt.plot(std_med['date'],std_med['cost_mean'],lw=2, alpha = 1,label='std_med')
plt.fill_between(std_med['date'], std_med['low_CI'], std_med['upper_CI'], alpha = 0.3,label='std_med_CI')
plt.plot(e7_med_72['date'],e7_med_72['cost_mean'],lw=2, alpha = 1,label='e7_med')
plt.fill_between(e7_med_72['date'], e7_med_72['low_CI'], e7_med_72['upper_CI'], alpha = 0.3,label='e7_med_CI')
plt.title('STD vs E7, med72')
plt.legend(loc='best',fontsize=10)
plt.savefig('med72.jpg')
In [18]:
plt.figure(figsize=(15,5),dpi=300)
plt.plot(std_hig['date'],std_hig['cost_mean'],lw=2, alpha = 1,label='std_hig')
plt.fill_between(std_hig['date'], std_hig['low_CI'], std_hig['upper_CI'], alpha = 0.3,label='std_hig_CI')
plt.plot(e7_hig_72['date'],e7_hig_72['cost_mean'],lw=2, alpha = 1,label='e7_hig')
plt.fill_between(e7_hig_72['date'], e7_hig_72['low_CI'], e7_hig_72['upper_CI'], alpha = 0.3,label='e7_hig_CI')
plt.title('STD vs E7, hig72')
plt.legend(loc='best',fontsize=10)
plt.savefig('hig72.jpg')
0.75/0.25
In [19]:
time_e7_75 = pd.read_csv('time_e7_75.csv')
In [20]:
e7_low_75 = time_e7_75[time_e7_75['usage']=='low'][['date','cost_mean']].groupby('date').mean()
e7_med_75 = time_e7_75[time_e7_75['usage']=='med'][['date','cost_mean']].groupby('date').mean()
e7_hig_75 = time_e7_75[time_e7_75['usage']=='hig'][['date','cost_mean']].groupby('date').mean()
e7_low_75.index=pd.to_datetime(e7_low_75.index)
e7_hig_75.index=pd.to_datetime(e7_hig_75.index)
e7_med_75.index=pd.to_datetime(e7_med_75.index)
In [22]:
e7_low_75.reset_index(inplace=True)
e7_med_75.reset_index(inplace=True)
e7_hig_75.reset_index(inplace=True)
#Create confidence interval
e7_low_75['low_CI'] = e7_low_75['cost_mean'] - np.std(e7_low_75['cost_mean'])
e7_low_75['upper_CI'] = e7_low_75['cost_mean'] + np.std(e7_low_75['cost_mean'])
e7_med_75['low_CI'] = e7_med_75['cost_mean'] - np.std(e7_med_75['cost_mean'])
e7_med_75['upper_CI'] = e7_med_75['cost_mean'] + np.std(e7_med_75['cost_mean'])
e7_hig_75['low_CI'] = e7_hig_75['cost_mean'] - np.std(e7_hig_75['cost_mean'])
e7_hig_75['upper_CI'] = e7_hig_75['cost_mean'] + np.std(e7_hig_75['cost_mean'])
e7_low_75 = e7_low_75[e7_low_75['date']>='20120201']
e7_med_75 = e7_med_75[e7_med_75['date']>='20120201']
e7_hig_75 = e7_hig_75[e7_hig_75['date']>='20120201']
In [23]:
plt.figure(figsize=(15,5),dpi=300)
plt.plot(std_low['date'],std_low['cost_mean'],lw=2, alpha = 1,label='std_low')
plt.fill_between(std_low['date'], std_low['low_CI'], std_low['upper_CI'], alpha = 0.3,label='std_low_CI')
plt.plot(e7_low_75['date'],e7_low_75['cost_mean'],lw=2, alpha = 1,label='e7_low')
plt.fill_between(e7_low_75['date'], e7_low_75['low_CI'], e7_low_75['upper_CI'], alpha = 0.3,label='e7_low_CI')
plt.title('STD vs E7, low75')
plt.legend(loc='best',fontsize=10)
plt.savefig('low_75.jpg')
In [25]:
plt.figure(figsize=(15,5),dpi=300)
plt.plot(std_med['date'],std_med['cost_mean'],lw=2, alpha = 1,label='std_med')
plt.fill_between(std_med['date'], std_med['low_CI'], std_med['upper_CI'], alpha = 0.3,label='std_med_CI')
plt.plot(e7_med_75['date'],e7_med_75['cost_mean'],lw=2, alpha = 1,label='e7_med')
plt.fill_between(e7_med_75['date'], e7_med_75['low_CI'], e7_med_75['upper_CI'], alpha = 0.3,label='e7_med_CI')
plt.title('STD vs E7, med75')
plt.legend(loc='best',fontsize=10)
plt.savefig('med75.jpg')
In [26]:
plt.figure(figsize=(15,5),dpi=300)
plt.plot(std_hig['date'],std_hig['cost_mean'],lw=2, alpha = 1,label='std_hig')
plt.fill_between(std_hig['date'], std_hig['low_CI'], std_hig['upper_CI'], alpha = 0.3,label='std_hig_CI')
plt.plot(e7_hig_75['date'],e7_hig_75['cost_mean'],lw=2, alpha = 1,label='e7_hig')
plt.fill_between(e7_hig_75['date'], e7_hig_75['low_CI'], e7_hig_75['upper_CI'], alpha = 0.3,label='e7_hig_CI')
plt.title('STD vs E7, hig75')
plt.legend(loc='best',fontsize=10)
plt.savefig('hig75.jpg')
0.8/0.2
In [27]:
time_e7_80 = pd.read_csv('time_e7_80.csv')
e7_low_80 = time_e7_80[time_e7_80['usage']=='low'][['date','cost_mean']].groupby('date').mean()
e7_med_80 = time_e7_80[time_e7_80['usage']=='med'][['date','cost_mean']].groupby('date').mean()
e7_hig_80 = time_e7_80[time_e7_80['usage']=='hig'][['date','cost_mean']].groupby('date').mean()
In [28]:
e7_low_80.index=pd.to_datetime(e7_low_80.index)
e7_hig_80.index=pd.to_datetime(e7_hig_80.index)
e7_med_80.index=pd.to_datetime(e7_med_80.index)
In [29]:
e7_low_80.reset_index(inplace=True)
e7_med_80.reset_index(inplace=True)
e7_hig_80.reset_index(inplace=True)
#Create confidence interval
e7_low_80['low_CI'] = e7_low_80['cost_mean'] - np.std(e7_low_80['cost_mean'])
e7_low_80['upper_CI'] = e7_low_80['cost_mean'] + np.std(e7_low_80['cost_mean'])
e7_med_80['low_CI'] = e7_med_80['cost_mean'] - np.std(e7_med_80['cost_mean'])
e7_med_80['upper_CI'] = e7_med_80['cost_mean'] + np.std(e7_med_80['cost_mean'])
e7_hig_80['low_CI'] = e7_hig_80['cost_mean'] - np.std(e7_hig_80['cost_mean'])
e7_hig_80['upper_CI'] = e7_hig_80['cost_mean'] + np.std(e7_hig_80['cost_mean'])
e7_low_80 = e7_low_80[e7_low_80['date']>='20120201']
e7_med_80 = e7_med_80[e7_med_80['date']>='20120201']
e7_hig_80 = e7_hig_80[e7_hig_80['date']>='20120201']
In [30]:
plt.figure(figsize=(15,5),dpi=300)
plt.plot(std_low['date'],std_low['cost_mean'],lw=2, alpha = 1,label='std_low')
plt.fill_between(std_low['date'], std_low['low_CI'], std_low['upper_CI'], alpha = 0.3,label='std_low_CI')
plt.plot(e7_low_80['date'],e7_low_80['cost_mean'],lw=2, alpha = 1,label='e7_low')
plt.fill_between(e7_low_80['date'], e7_low_80['low_CI'], e7_low_80['upper_CI'], alpha = 0.3,label='e7_low_CI')
plt.title('STD vs E7, low80')
plt.legend(loc='best',fontsize=10)
plt.savefig('low_80.jpg')
In [32]:
plt.figure(figsize=(15,5),dpi=300)
plt.plot(std_med['date'],std_med['cost_mean'],lw=2, alpha = 1,label='std_med')
plt.fill_between(std_med['date'], std_med['low_CI'], std_med['upper_CI'], alpha = 0.3,label='std_med_CI')
plt.plot(e7_med_80['date'],e7_med_80['cost_mean'],lw=2, alpha = 1,label='e7_med')
plt.fill_between(e7_med_80['date'], e7_med_80['low_CI'], e7_med_80['upper_CI'], alpha = 0.3,label='e7_med_CI')
plt.title('STD vs E7, med80')
plt.legend(loc='best',fontsize=10)
plt.savefig('med80.jpg')
In [34]:
plt.figure(figsize=(15,5),dpi=300)
plt.plot(std_hig['date'],std_hig['cost_mean'],lw=2, alpha = 1,label='std_hig')
plt.fill_between(std_hig['date'], std_hig['low_CI'], std_hig['upper_CI'], alpha = 0.3,label='std_hig_CI')
plt.plot(e7_hig_80['date'],e7_hig_80['cost_mean'],lw=2, alpha = 1,label='e7_hig')
plt.fill_between(e7_hig_80['date'], e7_hig_80['low_CI'], e7_hig_80['upper_CI'], alpha = 0.3,label='e7_hig_CI')
plt.title('STD vs E7, hig80')
plt.legend(loc='best',fontsize=10)
plt.savefig('hig80.jpg')
Comparation for the future price¶
training set and testing set to test which model is better
choose the arima to make forecast
Arima models¶
In [49]:
import pandas as pd
import numpy as np
import datetime as dt
from statsmodels.tsa.seasonal import seasonal_decompose
from matplotlib.pylab import style
from statsmodels.tsa.arima_model import ARIMA
from statsmodels.graphics.tsaplots import plot_acf, plot_pacf
from sklearn.metrics import mean_squared_error as MSE
from sklearn.metrics import mean_absolute_error as MAE
from sklearn.linear_model import LinearRegression
import matplotlib.pyplot as plt
import seaborn as sns
from matplotlib import style
plt.rcParams['savefig.dpi'] = 500
plt.rcParams['figure.dpi'] = 500
style.use('seaborn')
import warnings; warnings.simplefilter('ignore')
%matplotlib inline
In [69]:
std_train = std_med.copy()
std_train = std_train[['date','cost_mean']]
std_train.index = std_train.date
del std_train['date']
In [70]:
# choose the data from '20180101' to make predictions
std_train = std_train[std_train.index>'20180101']
std_train.plot(figsize=(15,5))
Out[70]:
In [71]:
# decomposition
decomposition =seasonal_decompose(std_train,model='mul')
std_train.trend = decomposition.trend
std_train.seasonal = decomposition.seasonal
std_train.residual = decomposition.resid
plt.figure(figsize=(15,9),dpi=200)
plt.subplot(311)
plt.plot(std_train.trend)
plt.title('trend')
plt.subplot(312)
plt.plot(std_train.seasonal)
plt.title('seasonal')
plt.subplot(313)
plt.plot(std_train.residual)
plt.title('residual')
Out[71]:
In [72]:
#first order difference
train_diff = std_train.diff()
train_diff = train_diff.dropna()
plt.figure(dpi=200,figsize=(15,5))
plt.plot(train_diff)
plt.title('first order diff')
Out[72]:
In [74]:
plot_acf(train_diff,lags=20)
Out[74]:
In [75]:
plot_pacf(train_diff,lags=20)
Out[75]:
In [76]:
# search for Optimal parameters p,d,q
import numpy as np
import itertools
p_min = 2
d_min = 1
q_min = 2
p_max = 8
d_max = 1
q_max = 8
# Initialize a DataFrame to store the results
results_bic = pd.DataFrame(index=['AR{}'.format(i) for i in range(p_min,p_max+1)],
columns=['MA{}'.format(i) for i in range(q_min,q_max+1)])
for p,d,q in itertools.product(range(p_min,p_max+1),
range(d_min,d_max+1),
range(q_min,q_max+1)):
if p==0 and d==0 and q==0:
results_bic.loc['AR{}'.format(p), 'MA{}'.format(q)] = np.nan
continue
try:
model = ARIMA(std_train, order=(p, d, q),
#enforce_stationarity=False,
#enforce_invertibility=False,
)
results = model.fit()
results_bic.loc['AR{}'.format(p), 'MA{}'.format(q)] = results.bic
except:
continue
results_bic = results_bic[results_bic.columns].astype(float)
In [77]:
fig, ax = plt.subplots(figsize=(10, 8),dpi=200)
ax = sns.heatmap(results_bic,
mask=results_bic.isnull(),
ax=ax,
annot=True,
fmt='.2f',
);
ax.set_title('BIC');
plt.savefig('ARIMA_std.png')
In [81]:
model_std = ARIMA(std_train, order=(2, 1, 2))
result_std = model_std.fit()
print(result_std.summary())
In [85]:
e772_train = e7_med_72[['date','cost_mean']]
e772_train.index = e772_train.date
e772_train = e7_med_72[['date','cost_mean']]
e772_train.index = e772_train.date
del e772_train['date']
del e772_train['date']
In [88]:
e772_train = e772_train[e772_train.index>'20180101']
e772_train.plot(figsize=(15,5))
Out[88]:
In [89]:
# decomposition
decomposition =seasonal_decompose(e772_train,model='mul')
e772_train.trend = decomposition.trend
e772_train.seasonal = decomposition.seasonal
e772_train.residual = decomposition.resid
plt.figure(figsize=(15,9),dpi=200)
plt.subplot(311)
plt.plot(e772_train.trend)
plt.title('trend')
plt.subplot(312)
plt.plot(e772_train.seasonal)
plt.title('seasonal')
plt.subplot(313)
plt.plot(e772_train.residual)
plt.title('residual')
Out[89]:
In [90]:
# search for Optimal parameters p,d,q
import numpy as np
import itertools
p_min = 2
d_min = 1
q_min = 2
p_max = 8
d_max = 1
q_max = 8
# Initialize a DataFrame to store the results
results_bic = pd.DataFrame(index=['AR{}'.format(i) for i in range(p_min,p_max+1)],
columns=['MA{}'.format(i) for i in range(q_min,q_max+1)])
for p,d,q in itertools.product(range(p_min,p_max+1),
range(d_min,d_max+1),
range(q_min,q_max+1)):
if p==0 and d==0 and q==0:
results_bic.loc['AR{}'.format(p), 'MA{}'.format(q)] = np.nan
continue
try:
model = ARIMA(e772_train, order=(p, d, q),
#enforce_stationarity=False,
#enforce_invertibility=False,
)
results = model.fit()
results_bic.loc['AR{}'.format(p), 'MA{}'.format(q)] = results.bic
except:
continue
results_bic = results_bic[results_bic.columns].astype(float)
In [91]:
fig, ax = plt.subplots(figsize=(10, 8),dpi=200)
ax = sns.heatmap(results_bic,
mask=results_bic.isnull(),
ax=ax,
annot=True,
fmt='.2f',
);
ax.set_title('BIC');
plt.savefig('ARIMA_e772.png')
In [92]:
model_e772 = ARIMA(e772_train, order=(2, 1, 2))
result_e772 = model_e772.fit()
print(result_std.summary())
In [93]:
e775_train = e7_med_75[['date','cost_mean']]
e775_train.index = e775_train.date
e775_train = e7_med_75[['date','cost_mean']]
e775_train.index = e775_train.date
del e775_train['date']
In [94]:
e775_train = e775_train[e775_train.index>'20180101']
e775_train.plot(figsize=(15,5))
Out[94]:
In [95]:
# decomposition
decomposition =seasonal_decompose(e775_train,model='mul')
e775_train.trend = decomposition.trend
e775_train.seasonal = decomposition.seasonal
e775_train.residual = decomposition.resid
plt.figure(figsize=(15,9),dpi=200)
plt.subplot(311)
plt.plot(e775_train.trend)
plt.title('trend')
plt.subplot(312)
plt.plot(e775_train.seasonal)
plt.title('seasonal')
plt.subplot(313)
plt.plot(e775_train.residual)
plt.title('residual')
Out[95]:
In [96]:
# search for Optimal parameters p,d,q
import numpy as np
import itertools
p_min = 2
d_min = 1
q_min = 2
p_max = 8
d_max = 1
q_max = 8
# Initialize a DataFrame to store the results
results_bic = pd.DataFrame(index=['AR{}'.format(i) for i in range(p_min,p_max+1)],
columns=['MA{}'.format(i) for i in range(q_min,q_max+1)])
for p,d,q in itertools.product(range(p_min,p_max+1),
range(d_min,d_max+1),
range(q_min,q_max+1)):
if p==0 and d==0 and q==0:
results_bic.loc['AR{}'.format(p), 'MA{}'.format(q)] = np.nan
continue
try:
model = ARIMA(e775_train, order=(p, d, q),
#enforce_stationarity=False,
#enforce_invertibility=False,
)
results = model.fit()
results_bic.loc['AR{}'.format(p), 'MA{}'.format(q)] = results.bic
except:
continue
results_bic = results_bic[results_bic.columns].astype(float)
In [97]:
fig, ax = plt.subplots(figsize=(10, 8),dpi=200)
ax = sns.heatmap(results_bic,
mask=results_bic.isnull(),
ax=ax,
annot=True,
fmt='.2f',
);
ax.set_title('BIC');
plt.savefig('ARIMA_e775.png')
In [98]:
model_e775 = ARIMA(e775_train, order=(2, 1, 2))
result_e775 = model_e775.fit()
print(result_std.summary())
In [108]:
pred_index = pd.date_range(start='2019-07-18', periods=60)
In [109]:
# Build Model
fitted_std = model_std.fit(disp=-1)
fitted_e775 = model_e775.fit(disp=-1)
# Forecast
fc_std, se, conf = fitted_std.forecast(60, alpha=0.05) # 95% conf
fc_e775, se_e775, conf_e775 = fitted_e775.forecast(60, alpha=0.05) # 95% conf
# Plot
plt.figure(figsize=(15,5), dpi=200)
plt.plot(std_train, label='training_std')
fc_series_std = pd.Series(fc_std, index=pred_index)
plt.plot(e775_train, label='training_e775')
fc_series_e775 = pd.Series(fc_e775, index=pred_index)
lower_series_std = pd.Series(conf[:, 0], index=pred_index)
upper_series_std = pd.Series(conf[:, 1], index=pred_index)
lower_series_e775 = pd.Series(conf_e775[:, 0], index=pred_index)
upper_series_e775 = pd.Series(conf_e775[:, 1], index=pred_index)
plt.plot(fc_series_std, label='forecast_std')
plt.fill_between(lower_series_std.index, lower_series_std, upper_series_std,
color='r', alpha=.15)
plt.plot(fc_series_e775, label='forecast_e775')
plt.fill_between(lower_series_e775.index, lower_series_e775, upper_series_e775,
color='b', alpha=.15)
plt.title('Forecast vs Actuals, 75%')
plt.legend(loc='upper left', fontsize=8)
Out[109]:
In [111]:
# Build Model
fitted_std = model_std.fit(disp=-1)
fitted_e772 = model_e772.fit(disp=-1)
# Forecast
fc_std, se, conf = fitted_std.forecast(60, alpha=0.05) # 95% conf
fc_e772, se_e772, conf_e772 = fitted_e772.forecast(60, alpha=0.05) # 95% conf
# Plot
plt.figure(figsize=(15,5), dpi=100)
plt.plot(std_train, label='training_std')
fc_series_std = pd.Series(fc_std, index=pred_index)
plt.plot(e772_train, label='training_e772')
fc_series_e772 = pd.Series(fc_e772, index=pred_index)
lower_series_std = pd.Series(conf[:, 0], index=pred_index)
upper_series_std = pd.Series(conf[:, 1], index=pred_index)
lower_series_e772 = pd.Series(conf_e772[:, 0], index=pred_index)
upper_series_e772 = pd.Series(conf_e772[:, 1], index=pred_index)
plt.plot(fc_series_std, label='forecast_std')
plt.fill_between(lower_series_std.index, lower_series_std, upper_series_std,
color='r', alpha=.15)
plt.plot(fc_series_e772, label='forecast_e772')
plt.fill_between(lower_series_e772.index, lower_series_e772, upper_series_e772,
color='b', alpha=.15)
plt.title('Forecast vs Actuals, 72%')
plt.legend(loc='upper left', fontsize=8)
Out[111]:
